Pith. sign in

REVIEW 3 major objections 7 minor 18 references

The physics residual of a neural CFD prediction is a reliable trust signal that shows where the field is wrong, but a poor objective for fixing it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 12:33 UTC pith:QLLQ7XA6

load-bearing objection Clean empirical dissociation: residual is a solid case-level trust signal for neural CFD, a bad fixer, with honest ablations and a residual-floor argument that mostly holds up. the 3 major comments →

arxiv 2607.10333 v1 pith:QLLQ7XA6 submitted 2026-07-11 physics.flu-dyn physics.comp-ph

NeuroForge: A Self-Correcting, Geometry-Native Neural CFD Engine with Calibrated Physics-Residual Trust

Ali Jabbary (Urmia University) , Kasra Ghanavati (University of Greenwich) This is my paper
classification physics.flu-dyn physics.comp-ph
keywords neural operatorssurrogate CFDphysics residualsuncertainty quantificationconformal predictiontrust calibrationairfoil aerodynamics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Machine-learning surrogates can predict steady flow fields over engineering geometries in milliseconds, but they emit a single field with no built-in way to know whether to trust it—especially on geometries far from the training set. This paper closes the loop with the governing physics by computing the discretised steady-RANS residual of the predicted field, a quantity available without ground truth, and tests what jobs that residual can actually do. The central finding is a clean two-way split: the residual is a reliable, backbone-robust case-level error detector whose rank correlation with field error stays positive across three architecturally different models and a second laminar dataset, and a split-conformal layer built on it attains target coverage; yet the same residual fails as a correction objective or acceptance gate, because residual and field error can move in opposite directions. A supervised deep-equilibrium corrector trained toward ground truth still cuts volume-field error on a strong backbone, but a controlled ablation shows the gain comes from the learned correction, not from feeding it the residual. The practical contribution is a self-auditing trust layer that tells a designer where not to trust a surrogate prediction, without claiming that residual minimisation will repair the field.

Core claim

The discretised steady-RANS residual of a neural-CFD prediction is a backbone-robust case-level trust signal—it ranks which whole predictions to distrust—with residual–error Spearman correlations of 0.625 on Transolver, roughly 0.40–0.83 on a grid Geo-FNO, 0.851 on MeshGraphNet, and 0.77 on DeepCFD laminar bluff bodies. The same residual is a poor correction objective: iteration sweeps raise the PDE residual while lowering field error, and residual-reducing acceptance gates accept almost nothing. A supervised DEQ corrector still reduces volume MSE by 9–25% on the strong backbone; zeroing its residual input matches that gain, so residual tells where error is, not how to remove it.

What carries the argument

The residual-floor theorem and the residual–error expansion: the monitored discrete residual (continuity and momentum, no-slip omitted) is nonzero at the discrete ground-truth field, so its minimiser is displaced from truth; far from truth the residual tracks error (trust signal), near truth the floor dominates and residual-reducing steps can increase field error (failed fixer).

Load-bearing premise

The residual-floor argument and residual-as-objective failure rest on the monitored discrete residual being substantially nonzero even at the discrete ground-truth field, partly because of coarse grid resolution and omitted near-wall terms.

What would settle it

On a wall-resolving residual operator and fine mesh where the residual of the ground-truth field is near zero, show residual-reducing correction iterations also reduce field error rather than raise residual while error falls; or find a backbone and dataset where case-level residual–error Spearman is consistently non-positive.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • CFD surrogates can ship a data-free case-level trust score and conformal uncertainty band without needing ground truth at inference time.
  • Residual-minimising correctors and residual-gated acceptance loops should not be assumed to reduce field error on steady RANS surrogates.
  • Supervised learned correctors can still improve strong backbones when trained toward truth rather than residual norms.
  • The residual trust signal generalises across neural architectures and at least one other flow regime (laminar bluff bodies).
  • Out-of-distribution coverage can hold empirically when uncertainty inflates under shift, even though the formal exchangeability guarantee does not transfer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The residual-as-trust pattern may transfer to other steady PDE surrogates (heat transfer, elasticity) where a discrete residual is cheap and ground truth is scarce during design exploration.
  • Wall-resolving residual operators that drive the ground-truth residual near zero could change whether residual minimisation remains a bad objective, testing the floor theorem’s load-bearing premise.
  • Early design loops that rank candidate geometries by residual magnitude could filter surrogate hallucinations before spending high-fidelity CFD budget.
  • Once classical fallback solvers are wired in, uncertainty-gated hybrid engines become practical: spend classical compute only where the residual trust layer flags low confidence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript studies what the discretised steady-RANS residual of a neural CFD surrogate can and cannot do. Its central claim is a two-way dissociation: the residual is a reliable, backbone-robust case-level trust signal (positive residual–error Spearman on Transolver 0.625±0.019, grid Geo-FNO ~0.40–0.83, MeshGraphNet 0.851±0.058, and DeepCFD ρ=0.77±0.12) but a poor correction objective or acceptance gate (iteration sweeps raise residual while field error falls; residual-reducing gates accept essentially nothing). A split-conformal layer attains target coverage (0.902±0.008 at 0.90), and with a deep-ensemble σ yields an input-adaptive band (ECE 0.074). Separately, a supervised DEQ corrector reduces volume MSE on a Transolver backbone (mse_u −9%, mse_v −21%, mse_p −25%); a residual-zero ablation attributes the gain to the learned correction, not residual conditioning. Section 6 supplies a residual-floor theorem explaining why a self-consistent monitor can detect error yet have a minimiser displaced from truth. Caveats (backbone-dependent correction, exchangeability for conformal guarantees) are stated explicitly.

Significance. If the dissociation holds, it is a practically useful result for neural CFD: a data-free residual can drive case-level trust and conformal bands without serving as a training or acceptance objective. Strengths that raise the contribution above a pure empirical note include (i) multi-seed, multi-architecture, and second-dataset evidence for the trust-signal half; (ii) a controlled residual-input ablation that cleanly attributes corrector gains; (iii) an explicit residual-floor theorem with real-data confirmation that ∥r⋆∥ is nonzero; (iv) honest reporting of backbone dependence, force self-integration limits, and vacuous acceptance-gate guarantees; and (v) a CPU-first open-source package with frozen I/O, committed scripts, and a hash manifest. The work is more diagnostic than architectural, but the diagnostic is load-bearing for how residual-based neural CFD pipelines should be designed.

major comments (3)
  1. [§5.5, Table 5] §5.5 and Table 5: the residual-as-objective negative is strongly supported by the iteration sweep (residual norm 0.11→0.62 while mse_u falls), but the companion claim that the backtracking acceptance test “accepts essentially none” is flagged by the authors themselves as the thinnest leg and lacks a committed accepted-step-count artifact. Because this claim is used to close the “objective/gate” half of the dissociation, please report quantitative acceptance rates (e.g., fraction of residual-reducing steps admitted over seeds/checkpoints) or demote the gate claim to a structural argument only.
  2. [§6, Theorem 1] §6, Theorem 1 (H1–H2): the residual-floor argument and the local-divergence cone rest on the monitored discrete operator Rh (continuity+momentum, no-slip omitted) having a substantially nonzero floor at the discrete ground truth, driven in part by 128² under-resolution and omitted ∇νt·∇u. The boundary-inclusive ρres sweep (results/control/bc_inclusive_sweep.json) is a useful robustness check and should be elevated into the main text, but it does not address resolution. A short sensitivity (or an explicit scoped claim) on whether the misplaced-minimum / local-divergence conclusions survive a wall-resolving residual definition would make the theorem’s domain of applicability precise without changing the empirical trust-signal half, which does not depend on H2.
  3. [Abstract; §3.3] §3.3 vs abstract/intro wording: the quantitative trust result is case-level (field-mean residual ranks whole predictions; per-cell Spearman only 0.22±0.06). Phrases such as “tells you where the prediction is wrong” and the traffic-light residual map can be read as validated spatial localisation. Please align abstract, thesis sentence, and §3.3 so that “where” unambiguously means which cases (or which coarse regions), and keep the per-cell map explicitly illustrative rather than certified.
minor comments (7)
  1. [Table 2, §5.2] Table 2 / deep-ensemble adaptive band: ECE 0.074 and q=2.35 are reported from a single 100/100 cal/test split. A brief multi-split or multi-seed note (even if only for coverage, not full ECE) would match the 3-seed standard used elsewhere.
  2. [§5.2, Tables 1–3, 7] Force metrics: the careful distinction between self-integrated field-to-field ρCl/ρCd and official OpenFOAM labels (§5.2) is excellent; ensure every table caption that reports ρCl/ρCd points to that caveat so readers do not compare to Bonnet et al. baselines in Table 7.
  3. [Figure 4] Figure 4 is the key “detector ≠ fixer” figure; the dual y-axes are dense. Consider a small inset or separate residual-vs-error scatter of the same sweep to make the opposite motion of residual and error unmistakable at a glance.
  4. [§2] Related work: Roy et al. (2025) and Song et al. (2026) are appropriately cited as residual-as-error-signal priors; a one-sentence contrast on steady RANS vs time-marching / epistemic-UQ settings would help non-specialist readers place the dissociation claim.
  5. [§3.3–3.4] Notation: residual magnitude is written ρres in §3.3 and residual_norm / N(y) elsewhere; a single symbol table or consistent notation would reduce scanning cost.
  6. [§7] The νt channel is near-degenerate (MSE ~5e−8) and residuals lean on laminar viscosity (§7). A short forward-looking sentence on whether a learned νt would change residual–error rank correlation would be useful.
  7. [Throughout] Minor polish: “arXiv:2607.10333v1” date line and a few long parenthetical result-path strings in the main text could move to footnotes or the supplement to improve readability.

Circularity Check

0 steps flagged

No significant circularity: trust-signal correlations, residual-as-objective negatives, residual-floor theorem, conformal coverage, and supervised corrector gains are all measured against held-out ground truth or derived from independently verified assumptions.

full rationale

The paper's central dissociation (residual as trust signal yes, correction objective no) is established by empirical measurements, not by construction. Residual–error Spearman correlations (Transolver 0.625±0.019, Geo-FNO ~0.40–0.83, MeshGraphNet 0.851±0.058, DeepCFD 0.77±0.12) compare the data-free residual of a prediction to field error against held-out ground-truth fields; nothing in the residual definition forces a positive rank correlation. The residual-as-objective negative is likewise empirical: iteration sweeps show residual norm rising (0.11→0.62) while mse_u falls (3.92→2.29). Theorem 1 (residual floor) derives misplaced minimizers from Rh(u∞)=0 and the independently verified H2 (Rh(u⋆)≠0 on 200 real AirfRANS cases, mean ||r⋆||=0.192); the decomposition Rh(û)=r⋆+Le is a first-order expansion of the monitored operator, not a fit of the conclusion. Conformal q is the standard split-conformal quantile on held-out nonconformity scores under a stated exchangeability assumption; coverage is then measured on a disjoint test split. The DEQ corrector is supervised toward ground truth, and the W1 ablation (residual channels zeroed at train and inference) attributes the −9/−21/−25% MSE gains to the learned correction rather than residual conditioning. No self-citations by the present authors appear in the reference list; all load-bearing citations (Transolver, AirfRANS, conformal prediction, DEQ, Banach contraction) are external. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled via self-citation. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The trust-signal claim is mostly empirical and rests on standard discrete RANS residual evaluation plus exchangeability for conformal guarantees. The residual-as-objective negative rests on the residual-floor axioms (monitored Rh definition and nonzero r⋆ at discrete truth). Free parameters are engineering knobs (trust fusion weights, DEQ κ, conformal α, grid size) that affect presentation and certificates but not the sign of residual–error correlation. No new physical entities are postulated.

free parameters (4)
  • trust fusion weights wr, wu = wr=0.6, wu=0.4
    Hand-chosen wr=0.6, wu=0.4 fuse residual and uncertainty into the traffic-light map T; not fitted to maximize residual–error correlation but affect the illustrative trust map.
  • DEQ contraction factor κ and damping = κ=0.9, damping=0.5
    Design parameters of the contractive corrector (κ=0.9, damping 0.5); contraction is verified empirically but κ is chosen, not derived from data.
  • conformal miscoverage level α and cal/test split size = α=0.1, n_cal=n_test=100
    α=0.1 and 100/100 splits set the coverage target and finite-sample quantile; standard conformal knobs, not physics constants.
  • Cartesian grid resolution (ny×nx) = 128×128 (certificate/ablation scale)
    128² rasterisation under-resolves high-Re boundary layers and contributes to nonzero residual floor at ground truth; a free discretisation choice that enters residual-floor H2.
axioms (5)
  • standard math Split-conformal coverage requires exchangeable calibration and test nonconformity scores (Proposition 1).
    Standard conformal prediction assumption; paper correctly states OOD coverage is empirical only once exchangeability fails.
  • domain assumption Monitored residual Rh is discrete steady incompressible RANS continuity+momentum on physical fields with νeff=ν+νt, omitting no-slip in residual_norm (H1).
    Defines the operator whose residual floor and trust signal are studied; BC-inclusive robustness is checked but the theorem is about this Rh.
  • domain assumption Residual floor is nonzero at discrete ground truth: r⋆:=Rh(u⋆)≠0 (H2).
    Load-bearing for Theorem 1 misplaced minimum and local divergence; empirically confirmed on AirfRANS test cases but partly a discretisation/closure artifact.
  • standard math Banach fixed-point applies when each layer of gθ is 1-Lipschitz and κ<1, so DEQ corrector contracts in δ (Proposition 2).
    Standard contraction mapping; verified with measured median factor 0.78. Governs δ, not PDE residual of the output field.
  • domain assumption Steady 2-D incompressible RANS with eddy viscosity is an adequate residual model for the AirfRANS regime studied.
    Background CFD modeling choice; νt channel is near-degenerate so residual leans on laminar viscosity.
invented entities (1)
  • NeuroForge residual-floor / residual-as-trust-vs-objective framing independent evidence
    purpose: Organizes why residual norm tracks error far from truth but its minimizer is displaced from u⋆ near truth.
    Conceptual packaging of a classical residual expansion plus operator-specific floor; not a new physical particle or force. Independent evidence is the empirical residual↑/error↓ sweep and nonzero ||r⋆|| measurements.

pith-pipeline@v1.1.0-grok45 · 31841 in / 3909 out tokens · 42600 ms · 2026-07-14T12:33:13.295633+00:00 · methodology

0 comments
read the original abstract

Machine-learning surrogates for computational fluid dynamics (CFD) predict steady flow fields orders of magnitude faster than classical solvers, but emit a single field with no built-in way to know whether to trust it -- especially out of distribution. We close the loop with the governing physics: we compute the discretised steady-RANS residual of the prediction and ask what jobs it can do. Our central finding is a two-way dissociation: the physics residual is a reliable, backbone-robust trust signal (it tells you where the prediction is wrong) but a poor correction objective (it does not tell you how to fix it). As a trust signal, the residual's per-case rank correlation with field error is consistently positive across three architecturally distinct backbones (Transolver 0.625+-0.019; grid Geo-FNO ~0.40, lifted to 0.83 by a learned corrector; MeshGraphNet 0.851+-0.058) and generalizes to a second dataset and flow regime (DeepCFD laminar bluff bodies, rho=0.77+-0.12). A split-conformal layer attains target coverage (0.902+-0.008 at the 0.90 target) and, paired with a deep-ensemble sigma, yields an input-adaptive band (ECE 0.074). As a correction objective or acceptance gate the residual fails: iteration sweeps raise the PDE residual while lowering field error. Alongside the trust layer we deploy a supervised deep-equilibrium corrector trained toward ground truth that reduces volume-field MSE on all three seeds (mse_u -9%, mse_v -21%, mse_p -25%) on the SOTA backbone; a controlled ablation zeroing the corrector's residual input matches it, so the gain is attributable to the learned correction, not residual-conditioning. We report caveats plainly: correction quality is backbone-dependent, and the coverage guarantee holds under exchangeability. The contribution is a self-auditing trust layer, the residual's two roles, and the learned self-correction it accompanies.

Figures

Figures reproduced from arXiv: 2607.10333 by Ali Jabbary (Urmia University), Kasra Ghanavati (University of Greenwich).

Figure 1
Figure 1. Figure 1: In-distribution ablation ( [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: OOD gap and the regime-invariant trust signal ( [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Honest matched-budget baseline; the grid backbone trails Transolver [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The money figure for the headline: PDE residual [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Conformal coverage stays in-band on OOD splits (empirical, via [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: In-distribution reliability / coverage-vs- [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The trust signal generalizes to a second dataset (DeepCFD, 2-D laminar flow over bluff bod [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-geometry OOD (qualitative): an airfoil-trained model on a cylinder. (a) predicted speed [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 1 canonical work pages

  1. [1]

    AhmedML: High-fidelity compu- tational fluid dynamics dataset for incompressible, low-speed bluff body aerodynamics.arXiv preprint arXiv:2407.20801,

    Neil Ashton, Danielle Maddix, Samuel Gundry, and Parisa Shabestari. AhmedML: High-fidelity compu- tational fluid dynamics dataset for incompressible, low-speed bluff body aerodynamics.arXiv preprint arXiv:2407.20801,

    Pith/arXiv arXiv

  2. [2]

    Florent Bonnet, Jocelyn Ahmed Mazari, Paola Cinnella, and Patrick Gallinari

    arXiv:1909.01377. Florent Bonnet, Jocelyn Ahmed Mazari, Paola Cinnella, and Patrick Gallinari. AirfRANS: High fidelity computational fluid dynamics dataset for approximating reynolds-averaged navier-stokes solutions. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track,

    Pith/arXiv arXiv 1909

  3. [3]

    Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed

    arXiv:2212.07564. Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed. DrivAerNet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks.arXiv preprint arXiv:2406.09624,

    Pith/arXiv arXiv

  4. [4]

    Yarin Gal and Zoubin Ghahramani

    arXiv:2103.12803. Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. InInternational Conference on Machine Learning (ICML),

    Pith/arXiv arXiv

  5. [5]

    Jun-Ting Hsieh, Shengjia Zhao, Stephan Eismann, Lucia Mirabella, and Stefano Ermon

    arXiv:1506.02142. Jun-Ting Hsieh, Shengjia Zhao, Stephan Eismann, Lucia Mirabella, and Stefano Ermon. Learning neu- ral PDE solvers with convergence guarantees. InInternational Conference on Learning Representations (ICLR),

    Pith/arXiv arXiv

  6. [6]

    Prashant K

    arXiv:1906.01200. Prashant K. Jha. Residual-based error corrector operator to enhance accuracy and reliability of neural operator surrogates of nonlinear variational boundary-value problems.Computer Methods in Applied Mechanics and Engineering, 419:116595,

    Pith/arXiv arXiv 1906

  7. [7]

    arXiv:2306.12047

    doi: 10.1016/j.cma.2023.116595. arXiv:2306.12047. Aditi S. Krishnapriyan, Amir Gholami, Shandian Zhe, Robert M. Kirby, and Michael W. Mahoney. Char- acterizing possible failure modes in physics-informed neural networks. InAdvances in Neural Information Processing Systems (NeurIPS),

    work page doi:10.1016/j.cma.2023.116595 2023

  8. [8]

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stu- art, and Anima Anandkumar

    arXiv:1612.01474. Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stu- art, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations (ICLR),

    Pith/arXiv arXiv

  9. [9]

    Zongyi Li, Daniel Z

    arXiv:2010.08895. Zongyi Li, Daniel Z. Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries.arXiv preprint arXiv:2207.05209,

    Pith/arXiv arXiv 2010

  10. [10]

    Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long

    arXiv:2308.05732. Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long. Transolver++: An accurate neural solver for PDEs on million-scale geometries.arXiv preprint arXiv:2502.02414,

    Pith/arXiv arXiv

  11. [11]

    Tanya Marwah, Ashwini Pokle, J

    arXiv:2402.01960. Tanya Marwah, Ashwini Pokle, J. Zico Kolter, Zachary C. Lipton, Jianfeng Lu, and Andrej Risteski. Deep equilibrium based neural operators for steady-state PDEs. InAdvances in Neural Information Processing Systems (NeurIPS),

    Pith/arXiv arXiv

  12. [12]

    Maziar Raissi, Paris Perdikaris, and George E

    arXiv:2312.00234. Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learn- ing framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707,

    Pith/arXiv arXiv

  13. [13]

    DoMINO: A decomposable multi-scale iterative neural operator for modeling large scale engineering simulations.arXiv preprint arXiv:2501.13350,

    Rishikesh Ranade, Mohammad Amin Nabian, Kaustubh Tangsali, Alexey Kamenev, Oliver Hennigh, Ram Cherukuri, and Sanjay Choudhry. DoMINO: A decomposable multi-scale iterative neural operator for modeling large scale engineering simulations.arXiv preprint arXiv:2501.13350,

    Pith/arXiv arXiv

  14. [14]

    Mateus Dias Ribeiro, Abdul Rehman, Sheraz Ahmed, and Andreas Dengel

    NVIDIA. Mateus Dias Ribeiro, Abdul Rehman, Sheraz Ahmed, and Andreas Dengel. DeepCFD: Efficient steady-state laminar flow approximation with deep convolutional neural networks.arXiv preprint arXiv:2004.08826,

    Pith/arXiv arXiv 2004

  15. [15]

    ANCHOR: Error-controlled adaptive numerical correction for neural operator time marching.arXiv preprint arXiv:2512.19643,

    Rajyasri Roy, Dibyajyoti Nayak, and Somdatta Goswami. ANCHOR: Error-controlled adaptive numerical correction for neural operator time marching.arXiv preprint arXiv:2512.19643,

    arXiv

  16. [16]

    Structure-aware epistemic uncertainty quantification for neural operator PDE surrogates.arXiv preprint arXiv:2603.11052,

    24 Haoze Song, Zhihao Li, Mengyi Deng, Xin Li, Duyi Pan, Zhilu Lai, and Wei Wang. Structure-aware epistemic uncertainty quantification for neural operator PDE surrogates.arXiv preprint arXiv:2603.11052,

    arXiv

  17. [17]

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long

    arXiv:2006.08591. Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for PDEs on general geometries. InInternational Conference on Machine Learning (ICML),

    Pith/arXiv arXiv 2006

  18. [18]

    arXiv:2402.02366. 25

    Pith/arXiv arXiv